Deep Learning Approach for Rapid Tsunami Height and Arrival Time Prediction

Elizabeth K. Su, & Lingsen Meng

Submitted September 7, 2025, SCEC Contribution #14649, 2025 SCEC Annual Meeting Poster #TBD

Accurate and timely tsunami warnings are critical for mitigating loss of life and property. Traditional deterministic simulations of tsunami waveforms using numerical methods such as finite element models can provide detailed predictions but are computationally intensive, making them less suitable for rapid response. For operational warning systems, however, complete waveform details are often unnecessary—what is essential are reliable estimates of maximum coastal wave heights and arrival times.

This study presents a hybrid modeling framework that leverages deep learning to significantly reduce prediction time while maintaining accuracy. We first generate training data using stochastic slip distributions for Mw 7.5–9.0 earthquakes along the Japan Trench, with corresponding maximum wave heights and arrival times computed via high-fidelity numerical simulations. A convolutional neural network enhanced with transformer blocks is then trained to learn the mapping from earthquake slip distributions to coastal tsunami metrics, capturing both local spatial patterns and long-range dependencies.

Once trained, the model can produce near-instant predictions, enabling probabilistic tsunami hazard assessment and rapid damage estimation. This approach demonstrates the potential of AI-driven methods to complement traditional physics-based modeling, accelerating the delivery of actionable tsunami warning information and improving disaster preparedness.

Key Words
tsunami early warning, machine learning

Citation
Su, E. K., & Meng, L. (2025, 09). Deep Learning Approach for Rapid Tsunami Height and Arrival Time Prediction. Poster Presentation at 2025 SCEC Annual Meeting.

Related Projects & Working Groups
Research Computing (RC)